In [1]:
import numpy as np
import pandas as pd
import matplotlib.pyplot as plt
import seaborn as sns
from sklearn.model_selection import train_test_split
from sklearn.preprocessing import StandardScaler
from sklearn.compose import ColumnTransformer
from sklearn.pipeline import Pipeline
from sklearn.ensemble import RandomForestClassifier
In [2]:
from scripts.feature_eng.feature_eng import process_data
from scripts.feature_eng.feature_eng import impute_cat_columns
from scripts.feature_eng.feature_eng import encode_cols
from scripts.model.utils import feature_selection
from scripts.model.model_params import param_grid
from scripts.model.utils import train_model, get_report, plot_precision_recall_curve, plot_roc_curve
In [3]:
random_state=2024
In [4]:
%matplotlib inline

Data Extraction¶

Load Data and Data dictionary¶

In [5]:
data = pd.read_excel('./data/Assessment.xlsx', sheet_name='Data')
data_dictionary = pd.read_excel('./data/Assessment.xlsx', sheet_name='Data Dictionary')

Data Transformation¶

  • Data Cleaning
  • Data visualization
  • Missing Value detection and imputation
  • Feature generation
  • Feature data type casting
  • Feature selection

Process the input data¶

In [6]:
data = process_data(data)

c_seg
NORMAL      55157
AFFLUENT    10926
Name: count, dtype: int64
In [7]:
x_cols = ['c_edu', 'c_hse', 'incm_typ', 'gn_occ', 'num_prd', 'casatd_cnt', 'mthcasa', 'maxcasa', 'mincasa', 'drvcr', 'mthtd', 'maxtd', 'asset_value', 'hl_tag', 'al_tag', 'pur_price_avg',
       'ut_ave', 'maxut', 'n_funds', 'cc_ave', 'max_mth_trn_amt', 'min_mth_trn_amt', 'avg_trn_amt', 'ann_trn_amt', 'ann_n_trx', 'cc_lmt','age_bin', 'wealth_accumulation', 
       'monthly_txn_frequency', 'credit_utilization', 'savings_behaviour', 'debt_to_asset_ratio', 'txn_freq_per_prd', 'investment_to_debt_ratio']
In [8]:
cat_cols = ['c_edu', 'c_hse', 'incm_typ', 'gn_occ', 'age_bin', 'hl_tag', 'al_tag', 'savings_behaviour']
In [9]:
cols_to_encode = ['c_edu', 'c_hse', 'gn_occ', 'age_bin']
num_cols = [i for i in x_cols if i not in cat_cols]
In [10]:
count_df = pd.DataFrame(data.groupby('c_id')['c_age'].count().reset_index())
count_df.rename(columns={'c_age': 'id_count'}, inplace=True)
duplicate_c_id = count_df[count_df.id_count >1]['c_id'].values.tolist()
non_duplicate_c_id = count_df[count_df.id_count ==1]['c_id'].values.tolist()
In [11]:
print(f"Number of unique customers: {count_df.shape[0]}")
print(f"% of customers with one record in the dataset {round(100*len(non_duplicate_c_id)/count_df.shape[0], 3)} %")
print(f"% of customers with multiple records for the same customer ID's {round(100*len(duplicate_c_id)/count_df.shape[0], 3)} %")

Number of unique customers: 47857
% of customers with one record in the dataset 69.94 %
% of customers with multiple records for the same customer ID's 30.06 %
In [12]:
non_duplicate_df = data[data.c_id.isin(non_duplicate_c_id)].copy()
duplicate_df = data[data.c_id.isin(duplicate_c_id)].copy()

Remove duplicate rows for customers who have multiple records with the same customer ID, keeping only one instance of the repeated rows.¶

In [13]:
duplicate_df.drop_duplicates(inplace=True)
In [14]:
data = pd.concat([non_duplicate_df, duplicate_df])
data.reset_index(inplace=True, drop=True)

Impute categorical columns: Replace the null value with the word "UNKNOWN"¶

In [15]:
data = impute_cat_columns(df=data, cat_columns=cat_cols)

Missing Values estimation & removal¶

In [16]:
missing_values_proportion = [(i,round(100*j/data.shape[0], 2)) for i,j in zip(data.columns, data.isnull().sum(axis=0).values) if j>0]
missing_val_df = pd.DataFrame(missing_values_proportion,columns=['Column_Name','Proportion'])
print(missing_val_df.to_markdown())
plt.figure(figsize =(15, 7))
sns.set_style('darkgrid')
ax=sns.barplot(x='Column_Name',y='Proportion',data=missing_val_df, palette = 'coolwarm')
plt.xticks(rotation=70)
ax.set(xlabel='Column Name', ylabel='Percentage Missing (%)',title='Missing value proportions per feature')
for i in ax.containers:
    ax.bar_label(i,)
plt.show()

|    | Column_Name              |   Proportion |
|---:|:-------------------------|-------------:|
|  0 | pc                       |         0.74 |
|  1 | casatd_cnt               |        37.97 |
|  2 | mthcasa                  |        40.92 |
|  3 | maxcasa                  |        40.92 |
|  4 | mincasa                  |        40.92 |
|  5 | drvcr                    |        55.04 |
|  6 | mthtd                    |        78.66 |
|  7 | maxtd                    |        78.66 |
|  8 | pur_price_avg            |        91.55 |
|  9 | ut_ave                   |        95.95 |
| 10 | maxut                    |        95.95 |
| 11 | n_funds                  |        95.95 |
| 12 | cc_ave                   |        74.16 |
| 13 | max_mth_trn_amt          |        82.25 |
| 14 | min_mth_trn_amt          |        82.25 |
| 15 | avg_trn_amt              |        82.25 |
| 16 | ann_trn_amt              |        82.25 |
| 17 | ann_n_trx                |        82.25 |
| 18 | cc_lmt                   |        27.75 |
| 19 | monthly_txn_frequency    |        82.25 |
| 20 | credit_utilization       |        74.17 |
| 21 | debt_to_asset_ratio      |        92.07 |
| 22 | txn_freq_per_prd         |        82.25 |
| 23 | investment_to_debt_ratio |        99.33 |

C:\Users\krish\AppData\Local\Temp\ipykernel_11772\1724794813.py:6: FutureWarning: 

Passing `palette` without assigning `hue` is deprecated and will be removed in v0.14.0. Assign the `x` variable to `hue` and set `legend=False` for the same effect.

  ax=sns.barplot(x='Column_Name',y='Proportion',data=missing_val_df, palette = 'coolwarm')
No description has been provided for this image
In [17]:
plot_cols = cat_cols.copy()
fig, axes = plt.subplots(nrows=len(plot_cols), ncols=1, figsize=(10, 8 * len(plot_cols)))

# Iterate through each categorical variable and create a bar plot
for i, var in enumerate(plot_cols):
    # Calculate proportions
    data_viz = data[data[var]!='UNKNOWN']
    proportions = data_viz.groupby(['c_seg', var]).size() / data_viz.groupby('c_seg').size()
    proportions = proportions.reset_index(name='Proportion')
    
    # Plot
    sns.barplot(x=var, y='Proportion', hue='c_seg', data=proportions, ax=axes[i])
    axes[i].set_title(f'Distribution of {var} by Customer Segment')
    axes[i].set_xlabel(var)
    axes[i].set_ylabel('Proportion')
    axes[i].legend(title='Customer Segment')
    axes[i].tick_params(axis='x', labelrotation=60, labelsize=10)

# Adjust layout and display the plot
plt.tight_layout()
plt.show()
No description has been provided for this image
In [18]:
data.groupby('c_seg')
target_proportions = data.groupby(['c_seg']).size() / data.shape[0]
target_proportions = target_proportions.reset_index(name='Proportion')
print(target_proportions.to_markdown())
fig, axes = plt.subplots(nrows=1, ncols=1, figsize=(10, 8))
sns.barplot(x='c_seg', y='Proportion', hue='c_seg', data=target_proportions, ax=axes)
axes.set_title(f'Distribution of label by Customer Segment')
axes.set_xlabel('c_seg')
axes.set_ylabel('Proportion')
axes.legend(title='Customer Segment')
axes.tick_params(axis='x', labelrotation=60, labelsize=10)

WARNING:matplotlib.legend:No artists with labels found to put in legend.  Note that artists whose label start with an underscore are ignored when legend() is called with no argument.

|    | c_seg    |   Proportion |
|---:|:---------|-------------:|
|  0 | AFFLUENT |     0.165338 |
|  1 | NORMAL   |     0.834662 |
No description has been provided for this image

Column processing and datatype casting¶

  • Replace the null value with zero
  • cast the categorical columns as category type
  • cast the numerical columns as float64 type
In [19]:
data[num_cols] = data[num_cols].fillna(value=0)

data[cat_cols] = data[cat_cols].astype('category')
data[num_cols] = data[num_cols].astype('float64')

Encode the following columns using LabelEncoder:¶

  • c_edu
  • c_hse
  • gn_occ
  • age_bin
In [20]:
data = encode_cols(df=data, cols_to_encode=cols_to_encode)

Pipeline for Numeric Data Transformation¶

This pipeline defines a preprocessing strategy for numeric data. It consists of a single step:

  1. StandardScaler Transformation: StandardScaler is applied to scale numeric features to have mean 0 and variance 1.
In [21]:
numeric_transformer = Pipeline(steps=[('scaler', StandardScaler())])
preprocessor = ColumnTransformer(transformers=[('num', numeric_transformer, num_cols)])

Split the dataset into random training and test sets, with 80% of the records allocated to the training set and 20% to the test set. Ensure that the split is stratified based on the target variable.¶

In [22]:
X_train, X_test, Y_train, Y_test = train_test_split(data[x_cols], data['target'], test_size=0.2, random_state=random_state, stratify=data['target'])
In [23]:
X_train.reset_index(inplace=True, drop=True)
X_test.reset_index(inplace=True, drop=True)


Y_train = Y_train.reset_index(drop=True)
Y_test = Y_test.reset_index(drop=True)

Feature Selection¶

In [24]:
clf = RandomForestClassifier(max_features='log2', n_jobs=-1, random_state=random_state, class_weight='balanced')
In [25]:
selector, selected_cols, selected_num_cols, selected_cat_cols = feature_selection(X=X_train[x_cols], 
                                                                                  Y=Y_train, 
                                                                                  cat_cols=cat_cols, 
                                                                                  num_cols=num_cols, 
                                                                                  clf=clf, 
                                                                                  scoring='f1', 
                                                                                  method='load')

Model Training & Evaluation¶

In [26]:
model_dict = {}
metrics_list = []
for model_name, model_attr in param_grid.items():
    print(f'Model Name: {model_name}')
    if model_name in ['LogisticRegression', 'MLPClassifier']:
        X_train_processed = preprocessor.fit_transform(X_train)
        X_train_processed = pd.DataFrame(X_train_processed, columns=num_cols)
        X_train_processed = pd.concat([X_train_processed, X_train[cat_cols]], axis=1)
        X_train_df = X_train_processed[selected_cols].copy()

        X_test_processed = preprocessor.fit_transform(X_test)
        X_test_processed = pd.DataFrame(X_test_processed, columns=num_cols)
        X_test_processed = pd.concat([X_test_processed, X_test[cat_cols]], axis=1)
        X_test_df = X_test_processed[selected_cols].copy()
    else:
        X_train_df = X_train[selected_cols].copy()
        X_test_df = X_test[selected_cols].copy()

    grid_search, best_estimator = train_model(X=X_train_df, 
                                              Y=Y_train, 
                                              model_name=model_name, 
                                              model=model_attr['model'], 
                                              model_param_grid=model_attr['param_grid'])
    predicted_probabilities = best_estimator.predict_proba(X_test_df)
    model_dict[f'{model_name}'] = {'model': best_estimator, 'search_meta': grid_search}

    print(''.join(200*['*']))
    plot_roc_curve(y_true=Y_test, predicted_prob=predicted_probabilities, model_name=model_name)
    plot_precision_recall_curve(y_true=Y_test, predicted_prob=predicted_probabilities, model_name=model_name)
    print(''.join(200*['*']))

    model_metrics_df = get_report(y_true=Y_test, predicted_prob=predicted_probabilities, model_name=model_name)
    metrics_list.append(model_metrics_df)

metrics_df = pd.concat(metrics_list)

Model Name: LogisticRegression

C:\Users\krish\.conda\envs\env_mb\lib\site-packages\sklearn\linear_model\_logistic.py:1216: UserWarning: 'n_jobs' > 1 does not have any effect when 'solver' is set to 'liblinear'. Got 'n_jobs' = 8.
  warnings.warn(

********************************************************************************************************************************************************************************************************
No description has been provided for this image 
********************************************************************************************************************************************************************************************************
Model Name: RandomForestClassifier
********************************************************************************************************************************************************************************************************
No description has been provided for this image 
********************************************************************************************************************************************************************************************************
Model Name: ExtraTreesClassifier
********************************************************************************************************************************************************************************************************
No description has been provided for this image 
********************************************************************************************************************************************************************************************************

Feature Importance¶

In [27]:
summary_list = metrics_df['summary'].values
model_list = metrics_df['Model_name'].values
threshold_list = metrics_df['Threshold'].values
cols = ['NORMAL', 'AFFLUENT']
In [28]:
result_temp = []
for summary, model_name, threshold in zip(summary_list, model_list, threshold_list):
    temp_df = pd.DataFrame(summary)[cols].T
    temp_df['threshold'] = threshold
    temp_df['Model Name'] = model_name
    result_temp.append(temp_df)
s_df = pd.concat(result_temp)
s_df.reset_index(inplace=True)
s_df.rename(columns={'index': 'class_d'}, inplace=True)
aff_s_df = s_df[s_df.class_d == 'AFFLUENT'].copy()
aff_s_df.reset_index(inplace=True, drop=True)
aff_s_df.sort_values('precision', ascending=False, inplace=True)
best_model = model_dict['RandomForestClassifier']['model']
In [29]:
feature_importance_df = pd.DataFrame()
feature_importance_df['feature'] =  selected_cols
feature_importance_df['importance'] =  best_model.feature_importances_
feature_importance_df.sort_values('importance',ascending=False, inplace=True)

Plot feature importance¶

In [30]:
plt.figure(figsize=(10, 6))
plt.barh(feature_importance_df['feature'], feature_importance_df['importance'], color='blue')
plt.xlabel('Importance')
plt.ylabel('Feature')
plt.title('Feature Importance')
plt.gca().invert_yaxis()  # Invert y-axis to have the highest importance at the top
plt.grid(False) 
plt.show()
No description has been provided for this image

Observation¶

In [31]:
# remove the outliers
plot_temp = data[data.wealth_accumulation < 50000].copy()
In [32]:
sns.set_style("whitegrid")
# Plot
plt.figure(figsize=(10, 6))
sns.scatterplot(x=plot_temp['wealth_accumulation'], y=np.arange(len(plot_temp)), hue=plot_temp['target'], palette='Set1', s=100)
plt.title('Wealth Accumulation vs Target Variable')
plt.xlabel('Wealth Accumulation')
plt.ylabel('Index')
plt.legend(title='Target')
plt.show()
No description has been provided for this image
In [33]:
# Plot
plot_temp_1 = plot_temp[plot_temp.wealth_accumulation >=1363.637012].copy()
plt.figure(figsize=(10, 6))
sns.scatterplot(x=plot_temp_1['wealth_accumulation'], y=np.arange(len(plot_temp_1)), hue=plot_temp_1['target'], palette='Set1', s=100)
plt.title('Wealth Accumulation vs Target Variable')
plt.xlabel('Wealth Accumulation')
plt.ylabel('Index')
plt.legend(title='Target')
plt.show()
No description has been provided for this image
In [34]:
plot_temp_1.groupby('c_seg').count()
Out[34]:
c_id c_age c_edu c_hse pc incm_typ gn_occ num_prd casatd_cnt mthcasa ... cc_lmt age_bin wealth_accumulation monthly_txn_frequency credit_utilization savings_behaviour debt_to_asset_ratio txn_freq_per_prd investment_to_debt_ratio target
c_seg
AFFLUENT 5460 5460 5460 5460 5217 5460 5460 5460 5460 5460 ... 5460 5460 5460 5460 5460 5460 5460 5460 5460 5460
NORMAL 3722 3722 3722 3722 3722 3722 3722 3722 3722 3722 ... 3722 3722 3722 3722 3722 3722 3722 3722 3722 3722

2 rows × 38 columns

In [35]:
plot_temp.groupby(['c_seg'])['wealth_accumulation'].describe()
Out[35]:
count mean std min 25% 50% 75% max
c_seg
AFFLUENT 10920.0 2074.191244 2831.931746 0.0 548.895433 1363.637012 2569.617043 45200.011250
NORMAL 55150.0 328.680022 1069.999620 0.0 0.000000 4.378062 167.217477 48312.248471